Regensburg 2022 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 2: Computational Materials Modelling: Energy Materials

MM 2.8: Talk

Monday, September 5, 2022, 12:15–12:30, H44

Adaptive kinetic Monte Carlo driven by local environment recognition — •King Chun Lai, Sebastian Matera, Christoph Scheurer, and Karsten Reuter — Fritz Haber Institute of the Max Planck Society, Berlin, Germany

Efficient lattice kinetic Monte Carlo (kMC) simulation generally relies on a complete prior-understanding of the possible elementary processes and their corresponding energy barriers. Adaptive kMC (akMC) overcomes this limitation by searching for those kinetics during the simulation, but the transition state searches (TSSs) become a bottleneck. To address this, we augment akMC with machine learning on local atomistic environments. Assigning a Smooth Overlap of Atomic Positions vector to each found process, we build up a database during simulation. This database is used to propose initial guesses for TSSs on basis of the proximity of the current local environments and the database entries. This proximity measure is self-adjusted based on statistics from TSSs on-the-fly. As the database fills, the proposed guesses from the database become close to the true transition states. These high quality TSS guesses improve simulation efficiency in two ways. First, the number of TSSs per kMC step gets significantly reduced by avoiding unsuccessful or repeating random TSSs. Second, damped Newton-Raphson becomes practical, which completes a proposed TSS in only a handful iterations. Taking a Pd surface as an illustrative example, we demonstrate the performance of the approach and also discuss how clustering and proximity learning can improve the TSS guesses further.